Pandiyan M, T Narendiranath Babu
School of Mechanical Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu, 631 014, India.
Sci Rep. 2025 May 25;15(1):18232. doi: 10.1038/s41598-025-01780-y.
Rolling element bearings (REBs) are crucial components in various industrial applications. The bearing faults occur due to prolonged operation, overloading, high speed, and inadequate lubrication. A bearing failure can lead to significant downtime and huge maintenance costs for the machines. Hence, industries require condition monitoring to reduce costs. This study presents an automated detection approach for diagnosing faults in REBs using a Customized Convolutional Neural Network (C-CNN). This work focuses on vibration signals sampled at 12,800 Hz and 5120 Hz as input data to perform the fault diagnosis of bearings. Further, the Multi Feature (MF) block has been used in the architecture of the C-CNN model for better accuracy. By incorporating techniques such as batch normalization and dropout, the model has improved stability and prevented overfitting. Using the Balance Cross-Entropy (BCE) loss function for training has helped the model optimize prediction accuracy by minimizing the difference between the actual class and the predicted class probabilities. A comparison with models such as MSCNN, CNN, RF, DBN, KNN, ANN, SVM, LSTM, ResNet and SqueezeNet was carried out. The proposed C-CNN model has been found to well perform other classifiers in accurately recognizing bearing faults, achieving an excellent accuracy of 95% and 93.5% on 12,800 and 5120 Hz datasets, respectively. Statistical significance tests and error bars validate the robustness of the model's performance. Extensive experiments were conducted to evaluate the impact of sampling frequency on diagnostic accuracy, hyperparameter tuning strategies, and model robustness under different noise levels and operating conditions. Furthermore, computational complexity analysis, including FLOPs estimation, was performed to assess real-time applicability. The findings indicate that the C-CNN approach is a reliable and efficient solution for bearing fault classification, offering significant practical implications for industrial condition monitoring systems and helping to prevent plant shutdown.
滚动轴承是各种工业应用中的关键部件。由于长时间运行、过载、高速运转以及润滑不足等原因，轴承会出现故障。轴承故障会导致机器大量停机时间和高昂的维护成本。因此，工业领域需要进行状态监测以降低成本。本研究提出了一种使用定制卷积神经网络（C-CNN）诊断滚动轴承故障的自动检测方法。这项工作重点关注以12800Hz和5120Hz采样的振动信号作为输入数据来进行轴承故障诊断。此外，在C-CNN模型架构中使用了多特征（MF）模块以提高准确性。通过纳入批量归一化和随机失活等技术，该模型提高了稳定性并防止了过拟合。使用平衡交叉熵（BCE）损失函数进行训练有助于模型通过最小化实际类别与预测类别概率之间的差异来优化预测准确性。与MSCNN、CNN、RF、DBN、KNN、ANN、SVM、LSTM、ResNet和SqueezeNet等模型进行了比较。结果发现，所提出的C-CNN模型在准确识别轴承故障方面比其他分类器表现更好，在12800Hz和5120Hz数据集上分别达到了95%和93.5%的优异准确率。统计显著性检验和误差线验证了模型性能的稳健性。进行了大量实验以评估采样频率对诊断准确性的影响、超参数调整策略以及不同噪声水平和运行条件下的模型稳健性。此外，还进行了计算复杂度分析，包括浮点运算次数（FLOPs）估计，以评估实时适用性。研究结果表明，C-CNN方法是一种可靠且高效的轴承故障分类解决方案，对工业状态监测系统具有重要的实际意义，并有助于防止工厂停产。
